It claims to be llama-compatible, so you can reuse the same ecosystem, like ollama for inference.
It will not run random code on your behalf and call OpenAI API like you fear.